Skip null PaperBench token usage during aggregation

[sylvesterkaczmarek]

Summary

Keep successful SimpleJudge runs from failing during post-grade token aggregation when a leaf has no OpenAI token-usage metadata.

SimpleJudge.grade_leaf() includes a token_usage key in judge_metadata, but generic/non-OpenAI completers can leave its value as None. The aggregation helper currently checks only whether the key exists and then passes the value to TokenUsage.from_dict(), which calls .items() and crashes on None.

Fixes #146.

Fix

Only deserialize leaf token usage when the metadata value is a dictionary. Null, missing, or otherwise unavailable usage contributes no token counts for that leaf.

Regression coverage

Adds tests verifying that:

• a leaf with {"token_usage": None} is skipped without error;
• normal usage from another leaf is still summed exactly;
• a leaf with no token-usage key produces an empty total rather than failing.

TokenUsage serialization, OpenAI usage field mapping, and valid usage summation are unchanged.